|
Geiger-Mode Avalanche Photon Diode (GM-APD) array Lidar can detect echo signals at the single photon level and obtain 3D distance information. However, due to the impact of ambient light and dark counting noise, the signal photon weak distance information is difficult to be extracted, and its imaging capability detrimentally decreases under low Signal-to Background Ratio (SBR). In this paper, we presents a deep learning framework specifically developed to extract single pixel distance values by employing classification techniques. The framework employs 1D convolutional neural network (1D CNN) coupled with a bidirectional long-short-term memory (BiLSTM) and proposes a new mechanism, distance sparse-attention mechanism (DSAM) to extract distance values from 3D point cloud data generated by GM-APD array LiDAR systems.Firstly, local features are extracted using 1D CNN, then the extracted feature sequences are fed into a bidirectional LSTM layer to capture global dependencies and weights are assigned to enhance the important features in combination with DSAM, and finally the prediction results are outputted by the fully connected layer. The accuracy of 95.2% is obtained on the sample test set, and the mean error of range measurement is about 0.0212 with a standard deviation of about 0.0331. The same LiDAR echo data collected during daytime is processed by using Deep Learning method, Peak Thresholding method, MLE and RJMCMC. The experimental results demonstrate the obvious advantages of the algorithm in this paper compared with the traditional algorithms under low SBR. This paper provides a new processing techniques for GM-APD LiDAR 3D distance imaging in complex environments, and establishes the basis for continuous, real-time monitoring.
|
